This repo contains the labels and scripts for the YOLO Solid Waste Assessment on Grids (SWAG) for detecting solid waste piles from UAV imagery. The model is based on a previous work, that can be found here.
YOLO SWAG is a pretrained YOLO26-cls model for semantic segmentation of solid waste piles in UAV (drone) imagery. The model classifies 5m x 5m grid cells as either "waste" or "background" based on polygon labels, enabling efficient waste detection in OpenAerialMap (OAM) scenes and other aerial datasets.
The GPKG-files in ./data/labels are following this naming schema: {continent}_{country}_{city}_{openaerialmap_id}_tiles.gpkg.
The original labels from the previous work are included aswell and follow the schema {openaerialmap_id}_tiles.gpkg
Download the aerial images from OpenAerialMap. After that run:
uv run python scripts/create_labeling_grid.py <path/to/oam_files> <path/to/output/directoy>The content of the output directory can be loaded into QGIS or similar applications for labelling. Tiles containing waste piles are labelled with class "1" and background with class "2".
The training scenes cover locations across the World Bank regions shown below.
The YOLO classifier was evaluated with 10 scene-held-out cross-validation folds. Each fold keeps all tiles from a scene in the same split, preventing spatial leakage between training, validation, and test data. Results are reported as mean ± standard deviation across folds.
| Split | Top-1 accuracy |
|---|---|
| Train | 92.60% ± 1.51% |
| Validation | 92.62% ± 1.53% |
| Test | 91.32% ± 3.16% |
The mean held-out-scene accuracy is 91.3%. The close training and validation performance suggests limited average overfitting, while the larger test variation indicates that some scenes are more challenging than others. Top-5 accuracy was 100% in every fold and is not informative for this task because there are fewer than five classes.
